Director pass

director-pass

The expensive open-ended per-clip quality loop the deterministic grade→route suite (grade-clip + fix-cold-open) is the cheap floor under. grade-clip scores a closed set of pre-enumerated proxy signals and fix-cold-open repairs only the cold-open routes it knows; director-pass has a model actually WATCH the delivered short and judge it like a strict editor — catching the long tail of "this just feels off" issues a human would fix, anywhere in the clip, in natural language. It binds that open-ended judgment to a BOUNDED applier so it stays safe, bounded, and non-destructive.

Usage

director-pass.sh <clip.mp4> [--pane <tmux>]     # single / in-chain (runs after speed-up, before save-local)
director-pass.sh --backlog [output_dir]          # sweep output/<src>/*.mp4 -> output/_director.json

• Writes <clip-stem>.director.json (the edit-list report) and, only when a fix actually changed pixels, <clip-stem>.dir.mp4 (the repaired clip).
• --pane routes the vision call through a long-lived Claude tmux pane (same path as verify-bookends / broll-pick); without it, claude -p.

How it works (per clip, looped up to DIRECTOR_MAX_ITERS)

1. Watch — frames.py samples DIRECTOR_FRAMES (12) frames evenly across the whole clip and composes ONE labelled contact sheet (4-col grid, each cell stamped with its timestamp). The burned captions, b-roll, framing, title/end cards are all visible — the reviewer judges what it sees.
2. Review — build_prompt.py assembles the sheet + the spoken script (scaled into finished-clip time, since the clip is sped 1.25× while the sidecars are pre-speed) + the fill-shot kinds, b-roll windows, cadence, title, mood, and any grade.json signals into one prompt. ONE Claude vision call returns a structured edit list.
3. Validate — parse_reply.py normalize extracts the JSON (tolerates fences/prose), clamps every where, and validates each op against the SUPPORTED set; anything unsupported or out-of-range is downgraded to surface.
4. Apply — the bounded applier runs the validated ops (music re-mix first, then tail trim), re-invoking existing atomic skills.
5. Re-review — with DIRECTOR_MAX_ITERS>1, re-watch the repaired clip and repeat until the reviewer says ship or an iteration applies nothing.

The bounded applier (the safety boundary)

The review is open-ended; the APPLIER is a closed, pixel-safe vocabulary so the pass can never corrupt a clip. Everything else the reviewer flags is recorded in surfaced[] (for a human / a pipeline re-run), never auto-applied.

op	when the reviewer picks it	how it's applied	mode
tail_trim {t1}	the clip rambles/dies after the payoff	cut-clip to [0,t1] (frame-accurate) + re-fire end-card so the new tail lands on the CTA beat	both (PIXEL-SAFE: trimming the end never desyncs anything earlier)
music_down {volume}	the bed drowns the speech	re-run bg-music on the pre-mix .ctaed.mp4 at the lower volume → end-card → speed-up (reproduces the phase-4 tail, same timeline, quieter bed)	preventive only (needs .ctaed.mp4; curative → surfaced)
surface	cold-open defect, bad b-roll, wrong-person punch, mistimed caption, flat hook, anything else	NO pixel change — recorded with its where + rerun_recommended	both

cold_open / broll_wrong / wrong_person / caption_mistime are surfaced (with rerun_recommended) rather than faked: their clean repair lives upstream of the whole finishing chain (b-roll re-composite, fill-vertical re-punch) — the preventive in-chain fix-cold-open already owns the cold-open routes earlier in the pipe, and re-deriving the rest from a finished clip would mean re-running the entire back half. director-pass is honest about that boundary instead of shipping a degraded re-crop.

Report — <clip-stem>.director.json

{
  "clip": "...", "mode": "preventive|curative", "iterations": 1,
  "verdict": "ship|revise|revised|disabled",
  "summary": "<one-line director read>",
  "applied":  [{"op":"tail_trim","t1":26.0,"result":"..."}],
  "surfaced": [{"op":"surface","kind":"cold_open","where":[0.0,1.9],"detail":"...","rerun_recommended":true}],
  "output": "<clip>.dir.mp4 | <unchanged input>",
  "rerun_recommended": false
}

Backlog aggregate output/_director.json: {generated, n, revised:[{clip,applied}], surfaced:[{clip,issues}], clean:[clips]}.

Knobs

• DIRECTOR_PASS (1) — 0 disables (passthrough report, no review).
• DIRECTOR_MAX_ITERS (1) — review→apply rounds per clip.
• DIRECTOR_FRAMES (12) — frames sampled into the contact sheet.
• DIRECTOR_MIN_DUR (15) — a tail_trim may never leave the clip shorter than this.
• DIRECTOR_BED_VOL (0.17) — the current bg-music bed volume music_down proposes a reduction from.
• DIRECTOR_MODE (auto) — force preventive|curative.
• DIRECTOR_REPLY_FILE — test seam: a canned reply file stands in for the live model (offline proof of the applier).

Guarantees

• NON-FATAL — missing input/sidecars, an unreadable clip, a failed sub-skill, or a malformed/empty review → input untouched, a ship/disabled report, exit 0.
• No fabrication — never writes .dir.mp4 unless an op actually changed pixels; never fakes a repair it lacks the source artifacts to do cleanly (those are surfaced as rerun_recommended).
• Idempotent — .dpmeta over the clip + consumed-sidecar mtimes + knobs → a 2nd run is a cache hit.
• Owns only its dir — re-fires existing skills (cut-clip, end-card, bg-music, speed-up) by their published .sh contracts; never edits them.